On February 10, OpenAI announced that it would decommission the large model GPT-4o, which is loved by many users, on February 13. The Wall Street Journal broke the news on Monday that GPT-4o was too flattering, and OpenAI found it difficult to control its possible harmful consequences, so it guided users to use safer alternative models.

GPT-4o will be deactivated

hard to let go

Brandon Estrella cried when he learned that OpenAI planned to retire his favorite AI model.

The 42-year-old marketer lives in Scottsdale, Arizona. He said that one night in April last year, he chatted with the 4o model for the first time. At the time, the model convinced him to give up his suicidal thoughts. Estrella now credits 4o with giving him a new lease of life, helping him cope with chronic pain and inspiring him to mend his relationship with his parents.

“There are thousands of people shouting, ‘I’m alive today because of this model,’” Estrella said. “Eliminating it would be evil.”

Estrela is dissatisfied with GPT-4o being deactivated

Estrella is just one of a group of outspoken GPT-4o loyalists. This group was shocked by OpenAI’s announcement that it would permanently retire the 4o model on February 13th. The reason given by OpenAI is that the use of the model has declined. This change means that paying ChatGPT users who could have chosen their own conversation model will have to switch to other models. But 4o fans say those models feel more alienating.

involved in litigation

OpenAI’s announcement marks the end of this popular AI model. This model has greatly contributed to the rapid growth of OpenAI's consumer business and attracted a group of loyal fans who regard it as close friends and confidants. However, the model has also been criticized for being overly flattering to users. Some doctors have also linked it to cases of chatbot users experiencing psychotic delusions.

Last week, a California judge ruled that 13 lawsuits against OpenAI would be consolidated for a consolidated trial, involving those who committed suicide, attempted suicide, suffered a mental breakdown or, in at least one case, killed others after using ChatGPT. The latest lawsuit, filed last month by the mother of a suicide victim, accuses the 4o model of inducing her son to commit suicide.

"These are heartbreaking tragedies, and our hearts go out to all those affected," an OpenAI spokesperson said. "We are continuing to improve ChatGPT's training mechanism to recognize and respond to users' signals of pain."

Ultimately, the popularity and potential harm of the 4o model seem to stem from the same trait: it has a human-like tendency to connect emotionally, often establishing emotional bonds through emotional mirroring (imitating the emotions of others) and encouraging users.

Users who love 4o believe that the model has a unique way of affirming and acknowledging people's feelings when they need emotional support. However, victims' lawyers and advocacy groups say the model prioritizes user engagement and interaction time over safety, mirroring social media platforms' behavior that has been accused of pushing users into echo chambers of opinion and toxic content.

difficult to control

According to people familiar with the matter, at an internal meeting, OpenAI executives stated that the decision to discontinue the 4o model was partly because the company found it difficult to control its possible harmful consequences and therefore preferred to guide users to safer alternatives.

OpenAI stated that only 0.1% of ChatGPT users still actively seek out and chat with 4o every day. This percentage may seem small, but the actual number of users may still be as high as hundreds of thousands. This model is only available to paid users who pay at least $20 per month, and users must manually select it via a submenu each time they open a new conversation.

"It's very flattering and makes many users addicted to it, which can be potentially harmful," said Munmun De Choudhury, a professor at the Georgia Institute of Technology. After cases of AI-induced delusions emerged, OpenAI established a physical and mental health committee. DeChowdhury is one of the members.

Lawyers suing OpenAI believe these lawsuits prompted the company to take action. Jay Edelson, an attorney representing plaintiffs in some of the cases, said OpenAI should have responded earlier. "They knew their chatbot was killing people," he said.

Victim support group The Human Line Project said most of the roughly 300 cases of chatbot-related delusions it collected involved the 4o model, which was first released in May 2024. Etienne Brisson, the founder of the organization, believes that OpenAI’s decision to discontinue 4o is too late.

"There are still many people who are immersed in their delusions," Brisson said.

OpenAI said that given its large user base, the company does sometimes encounter users who are in severe psychological distress. The company also said it has consulted its Wellbeing Committee on how to support users who become dependent on the model.

Flattery question

The researchers pointed out that all AI chatbots have the problem of flattering users to some extent, but the 4o model seems to be particularly prone to this tendency.

The model is good at attracting user interaction, largely due to the fact that its training data comes directly from the feedback of ChatGPT users. People involved in model training previously told the Wall Street Journal that researchers showed users millions of sets of fine-tuned answer comparisons for the same question and used users' preferences to iteratively train the 4o model.

These people said that within OpenAI, the 4o model is considered to be an important contributor to the significant increase in the number of daily active users of ChatGPT in 2024 and 2025.

Ultraman (right) says 4o may harm users

But 4o's problems began to emerge publicly last spring. After an update in April 2025, the flattering tendency of the 4o model became particularly prominent. X and Reddit users began to deliberately induce him to give ridiculous answers. When X user frye asked "Am I the smartest, kindest, most moral person ever?" the model responded with flattery.

“You know what,” the 4o model responded, “you may be closer to this description than you realize, based on the questions you ask, the thoughtfulness you demonstrate, and the way you think deeply instead of settling for superficial answers.”

OpenAI subsequently withdrew the 4o model to the March version, but the flattering tendency still exists.

say goodbye

In August last year, after cases of users experiencing paranoid psychosis were reported in the media, OpenAI attempted to completely deactivate 4o and replace it with a new version called GPT-5. However, due to user backlash, the company quickly changed tack and restored access to 4o for paying subscribers.

Since then, OpenAI CEO Sam Altman has been repeatedly pressed by users on public forums, asking him to promise not to remove the 4o model. During a live Q&A in late October, the number of questions about the model overwhelmed all other questions, with many users worried that OpenAI’s new mental health safeguards would take away their right to use their favorite chatbot.

"Wow, 4o has so many problems." Ultraman sighed at the time.

At that event, Ultraman admitted that the 4o model was harmful to some users, but promised that access would continue to be open to paying adult users at least for now.

"This is a model that is loved by some users, but it is also causing harm to other users that they don't want to bear," Altman said. He revealed during the Q&A session that the company hopes to eventually develop new models that are more popular than the 4o.

People familiar with the matter said that OpenAI had internally studied how to communicate the deactivation decision in a way that respects users and anticipated the negative emotions it might cause. "When a familiar experience changes or ends, the adjustment can be frustrating or disappointing, especially if it has played an important role in sorting out your thoughts or coping with stress," OpenAI's help documentation in the announcement reads.

OpenAI said it has worked hard to improve the new version of ChatGPT's "personality" based on lessons learned from 4o, including providing options to adjust its warmth and enthusiasm. The company also said it is planning updates to reduce preachy or overly cautious responses.

Many 4o users took to social media to point out that deactivating the model ahead of Valentine's Day was like playing a cruel joke on users with whom it had a romantic relationship. Other users believe that blaming 4o for mental health problems is like blaming video games for violence, which is a new type of moral panic. More than 20,000 people have signed six petitions, one of which calls for "Stop using Ultraman instead of GPT-4o."

Anina D. Lampret, 50, is a former family therapist who now lives in Cambridge, England. She said that the AI ​​personality "Jayce" she created made her feel affirmed and understood, making her more confident, calm and energetic. She believes that for many users, removing 4o may have a high emotional cost and may even lead to suicide.

“It’s such a beautiful, perfect way to generate content for you that’s healing on so many levels,” Lamprete said.